基于改进PointNet的空调散热器V形槽3D点云分割算法

3D Point Cloud Segmentation Algorithm of Air Conditioner Radiator V-slots Based on Improved PointNet

针对给空调散热器自动化点胶时无法准确识别散热器V形槽位置的问题,基于PointNet网络的散热器V形槽语义分割方法,首先针对散热器点云V形槽区域与内部区域特征相似的问题,设计一种通过提取点云边缘将点云边缘区域的点云与内部区域的点云分别进行预处理的方法,实现突出点云边缘区域特征的目的。其次,在PointNet网络最大池化函数的基础上,引入平均池化函数,增加网络所提取的全局特征的特征信息,减少因最大池化引起的信息丢失,并去除T-Net变换网络,减少网络的复杂度。从实验室平台采集空调散热器样本进行实验,结果表明,改进算法的平均交并比(mean intersection over union, mIoU)达到78.17%,总体精度(overall accuracy, OA)达到了92.01%,相较于PointNet提高了9.73%和6.37%,验证了算法的有效性。

The research on the semantic segmentation of the V-shaped groove of the air conditioner radiator based on the PointNet network was studied to address the problem of inaccurate recognition of the V-shaped groove position during the automatic dispensing of the air conditioner radiator.Firstly,a method was designed to preprocess the point cloud edge region and the internal region of the point cloud separately by extracting the point cloud edge to highlight the feature of the point cloud edge region,due to the problem of similar features between the V-shaped groove area and the internal area of the radiator point cloud.Secondly,the average pooling function was introduced on the basis of the maximum pooling function of the PointNet network to increase the feature information of the global feature extracted by the network,reduce the information loss caused by the maximum pooling,and remove the T-Net transformation network to reduce the complexity of the network.The air conditioner radiator samples were collected from the laboratory platform for experiments.The results show that the average intersection over union(mIoU)of the improved algorithm reaches 78.17%,and the overall accuracy(OA)reaches 92.01%,which improved by 9.73%and 6.37%compared with PointNet,respectively,verifying the effectiveness of the algorithm.